Best local AI models for NVIDIA NVS 510

The largest models that fit fully

The 30 largest of the 56 models that fit; every smaller model in the catalog fits too. Best quant means the highest quality compression whose weights and 4k context both sit inside the memory.

Close, but only with CPU offload

These need more than the card holds at their smallest practical quant, so part of the model runs from system memory (figures assume 32 GB of it). They work, several times slower.

How to read this

The NVIDIA NVS 510 is an entry level graphics card equipped with 2 GB of DDR3 memory. Because of this small memory capacity, running artificial intelligence models locally requires careful selection of model sizes and quantization levels. The memory size of the card determines the maximum size of the model that can reside entirely on the hardware during processing.

The quantization column indicates the specific compression format used to shrink the model. Quantization formats like Q4_K_M, Q5_K_M, Q6_K, and Q8_0 reduce the precision of the model weights. This reduction allows larger models to fit into the limited 2 GB memory space of the card. For example, the Allegro 2.8B model fits into 2 GB of used memory when using the Q4_K_M quantization. Similarly, the Open-Sora Plan 2.7B model fits into 2 GB of used memory at the Q4_K_M quantization.

Several other models can run entirely within the onboard memory. The LFM2 2.6B and Playground v2.5 2.6B models both use 1.9 GB of memory at Q4_K_M. The Stable Diffusion 3.5 Medium and Canary 2.5B models require 1.8 GB of memory at Q4_K_M. If you use higher precision quants like Q6_K, the SmolVLM 2B, Stable Diffusion 3 Medium, Pyramid Flow 2B, and Wav2Vec2 XLS-R models will use exactly 2 GB of memory. The Moondream 2 1.9B model uses 1.9 GB of memory at Q6_K, while the Qwen3 1.7B and SmolLM2 1.7B models use 1.7 GB of memory.

When a model exceeds the 2 GB onboard memory, you must use CPU offload. CPU offload splits the workload between the graphics card and your system RAM. This offload process comes with a performance cost because DDR3 memory and system RAM buses are much slower than dedicated graphics pipelines. For example, running the SmolLM3 3B, Replit Code v1.5 3B, Kandinsky 3.1, Voxtral Mini, Orpheus TTS, or Higgs Audio v2 requires 2.2 GB of VRAM at Q4_K_M and 4.2 GB of system RAM. The MusicGen 3.3B model requires 2.4 GB of VRAM and 4.4 GB of system RAM. Larger models like Stable Diffusion XL 3.417B require 4.1 GB of VRAM at FP8 and 6.1 GB of system RAM.

Users must also consider the context window when running these models. The memory usage figures listed are calculated at a base 4k context window. Expanding the context window beyond 4k tokens will significantly increase memory consumption. This extra memory usage can easily exceed the 2 GB limit of the card and force the system to use slower system RAM.